Skip to content

Add LabelEncoder CUDA execution provider for numeric types #28045

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
tianleiwu merged 12 commits into from
Apr 25, 2026
Merged

Add LabelEncoder CUDA execution provider for numeric types #28045

Show file tree
Hide file tree
Changes from 8 commits
Commits
Show all changes
12 commits
Select commit Hold shift + click to select a range
492f2b4
Initial plan
Copilot Apr 13, 2026
7686a23
Add CUDA LabelEncoder implementation with binary search lookup
Copilot Apr 13, 2026
d7ae553
Add unit tests for CUDA LabelEncoder including NaN key handling
Copilot Apr 13, 2026
728c396
Address code review: add comments for double type specializations and…
Copilot Apr 13, 2026
b58706f
Merge remote-tracking branch 'origin/main' into copilot/add-labelenco…
tianleiwu Apr 15, 2026
5285d96
lintrunner
tianleiwu Apr 15, 2026
c956075
fix plugin ep build
tianleiwu Apr 16, 2026
426d878
fix build
tianleiwu Apr 16, 2026
1c33ed5
Address review feedback: remove unused include, use SafeInt, add empt…
tianleiwu Apr 19, 2026
59ff360
fix(cuda): use stable_sort + dedup in LabelEncoder to match CPU seman…
tianleiwu Apr 20, 2026
24c59b1
fix(cuda): include LabelEncoder in minimal builds
tianleiwu Apr 22, 2026
5d3269b
refactor(cuda): extract tensor attribute helper
tianleiwu Apr 22, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
5 changes: 5 additions & 0 deletions docs/OperatorKernels.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1028,6 +1028,11 @@ Do not modify directly.*
|Xor|*in* A:**T**<br> *in* B:**T**<br> *out* C:**T1**|7+|**T** = tensor(bool)<br/> **T1** = tensor(bool)|
| |
| |
|**Operator Domain:** *ai.onnx.ml*||||
|LabelEncoder|*in* X:**T1**<br> *out* Y:**T2**|4+|**T1** = tensor(double), tensor(float), tensor(int64)<br/> **T2** = tensor(double), tensor(float), tensor(int64)|
Comment thread
tianleiwu marked this conversation as resolved.
|||[2, 3]|**T1** = tensor(float), tensor(int64)<br/> **T2** = tensor(float), tensor(int64)|
| |
| |
|**Operator Domain:** *com.microsoft*||||
|Attention|*in* input:**T**<br> *in* weights:**T**<br> *in* bias:**T**<br> *in* mask_index:**M**<br> *in* past:**T**<br> *in* attention_bias:**T**<br> *in* past_sequence_length:**M**<br> *out* output:**T**<br> *out* present:**T**|1+|**T** = tensor(bfloat16), tensor(float), tensor(float16)|
|BeamSearch|*in* input_ids:**F**<br> *in* max_length:**I**<br> *in* min_length:**I**<br> *in* num_beams:**I**<br> *in* num_return_sequences:**I**<br> *in* length_penalty:**T**<br> *in* repetition_penalty:**T**<br> *in* vocab_mask:**M**<br> *in* prefix_vocab_mask:**M**<br> *in* attention_mask:**I**<br> *in* decoder_input_ids:**I**<br> *in* logits_processor:**I**<br> *out* sequences:**I**<br> *out* sequences_scores:**T**<br> *out* scores:**T**|1+|**T** = tensor(float), tensor(float16)|
Expand Down
57 changes: 52 additions & 5 deletions onnxruntime/core/providers/cuda/cuda_execution_provider.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -564,6 +564,55 @@ Status CUDAExecutionProvider::ReplayGraph(int graph_annotation_id) {
}

namespace cuda {

template <>
KernelCreateInfo BuildKernelCreateInfo<void>() {
return {};
}

#ifndef DISABLE_ML_OPS
// LabelEncoder (ML domain) - numeric types only
class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kMLDomain, 2, 3, float_int64, LabelEncoder);
class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kMLDomain, 2, 3, int64_float, LabelEncoder);
class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kMLDomain, 2, 3, float_float, LabelEncoder);
class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kMLDomain, 2, 3, int64_int64, LabelEncoder);
class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kMLDomain, 4, int64_int64, LabelEncoder);
class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kMLDomain, 4, int64_float, LabelEncoder);
class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kMLDomain, 4, float_int64, LabelEncoder);
class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kMLDomain, 4, float_float, LabelEncoder);
class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kMLDomain, 4, double_double, LabelEncoder);
class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kMLDomain, 4, double_int64, LabelEncoder);
class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kMLDomain, 4, int64_double, LabelEncoder);

Status RegisterOnnxMLOperatorKernels(KernelRegistry& kernel_registry) {
static const BuildKernelCreateInfoFn function_table[] = {
BuildKernelCreateInfo<void>, // default entry to avoid the list become empty after ops-reducing

// LabelEncoder (ML domain) - numeric types only
BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kMLDomain, 2, 3, float_int64, LabelEncoder)>,
BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kMLDomain, 2, 3, int64_float, LabelEncoder)>,
BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kMLDomain, 2, 3, float_float, LabelEncoder)>,
BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kMLDomain, 2, 3, int64_int64, LabelEncoder)>,
BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kMLDomain, 4, int64_int64, LabelEncoder)>,
BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kMLDomain, 4, int64_float, LabelEncoder)>,
BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kMLDomain, 4, float_int64, LabelEncoder)>,
BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kMLDomain, 4, float_float, LabelEncoder)>,
BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kMLDomain, 4, double_double, LabelEncoder)>,
BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kMLDomain, 4, double_int64, LabelEncoder)>,
BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kMLDomain, 4, int64_double, LabelEncoder)>,
};

for (auto& function_table_entry : function_table) {
KernelCreateInfo info = function_table_entry();
if (info.kernel_def != nullptr) { // filter disabled entries where type is void
ORT_RETURN_IF_ERROR(kernel_registry.Register(std::move(info)));
}
}

return Status::OK();
}
#endif

// opset 1 to 9
class ONNX_OPERATOR_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, MemcpyFromHost);
class ONNX_OPERATOR_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, MemcpyToHost);
Expand Down Expand Up @@ -1765,11 +1814,6 @@ class ONNX_OPERATOR_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 25, T
class ONNX_OPERATOR_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 25, Unsqueeze);
#endif

template <>
KernelCreateInfo BuildKernelCreateInfo<void>() {
return {};
}

static Status RegisterCudaKernels(KernelRegistry& kernel_registry) {
static const BuildKernelCreateInfoFn function_table[] = {
BuildKernelCreateInfo<void>, // default entry to avoid the list become empty after ops-reducing
Expand Down Expand Up @@ -3009,6 +3053,9 @@ static std::shared_ptr<KernelRegistry> s_kernel_registry;
void InitializeRegistry() {
s_kernel_registry = KernelRegistry::Create();
ORT_THROW_IF_ERROR(cuda::RegisterCudaKernels(*s_kernel_registry));
#ifndef DISABLE_ML_OPS
ORT_THROW_IF_ERROR(cuda::RegisterOnnxMLOperatorKernels(*s_kernel_registry));
#endif
}

void DeleteRegistry() {
Expand Down
Loading
Loading